sceptre

EasyOCR's accuracy. Rust's speed and footprint.

A from-scratch Rust reimplementation of EasyOCR's OCR pipeline — CRAFT text detection then gen2 CRNN recognition with CTC decoding, over ONNX. It agrees with EasyOCR's output across eight scripts (English, Latin, Chinese-simplified, Japanese, Korean, Cyrillic, Telugu, Kannada) on the ort backend's CPU execution provider — the tract backend uses a fixed detection canvas that can group text lines differently (see ADR 0027) — with no Python runtime, and runs on native ONNX Runtime (ort) or a pure-Rust backend (tract) behind one seam.

Models download from Hugging Face on first use, cache locally, and are sha256-verified on download — every run after that reads the cache with no network.

Embedding hosts can instead supply registry-described ONNX bytes through VerifiedModelProvider; build_warmed verifies and initializes the detector and selected recognizer once.

Usage

use sceptre::{Reader, ReadOptions};

let reader = Reader::builder().build()?;
for line in reader.readtext("receipt.png".as_ref(), &ReadOptions::default())?.lines {
    println!("{} ({:.2})", line.text, line.confidence);
}
# Ok::<(), sceptre::OcrError>(())

The crate ships default = []; enable a backend and model download: sceptre = { version = "0.4", features = ["ort-bundled", "download"] }. ort-bundled fetches a prebuilt ONNX Runtime at build time; on targets ort publishes no prebuilt for — Intel macOS, musl/Alpine, armv7, riscv64, FreeBSD, i686, s390x, powerpc64le — use ort-dynamic (bring your own libonnxruntime) or tract (pure Rust) instead.

For the CLI, install the sceptre-cli crate; its default build is self-contained (ort-bundled + download).

Full documentation, benchmarks, and design notes live at the project repository.

License

MIT. Model weights are distributed by third parties under their own licenses (Apache-2.0).